in reply to Dmytri

mastodon - Link to source

sw

@criffer ah, we speak about AI tools to create code...

Your "spec" must be as good & as exact as possible.

We - in my area - have iterations/sprints/feedback cycles in order to close gaps, clarify misunderstandings, learn, etc with the target to have the product that you want - interaction between e.g. PO and Dev. For me that is one aspect of agile, scrum - speak to each other.

How such a communication can be executed between POs and AI-tools I cannot comment - not my working mode.

@Christina Jennifer
in reply to Stephan Lukasczyk

glitchsoc - Link to source

Dmytri

@stephan @ainmosni I'm actually pretty positive about LLMs and software engineering, when they are an interface that helps me code, so help with syntax, boilerplate, snippets, feedback, docs, etc. However I don't see autonomous Agents/Vibe coding is really viable for anything larger than a personal use tool, quick prototype/poc, or basic scaffolding for the reasons I've raised, and this can't be fixed by improving agents, because it's a consequence of reduced systemic learning that drives xp, agile, lean, etc. Which also means I'm much more excited about the smaller free models like Qwen Coder, as they are good enough for my use cases, so I see the real innovation being in small, specialized models and improved UX/DX.
@Stephan Lukasczyk @Daniël Franke
in reply to Dmytri

gotosocial - Link to source

Daniël Franke

I'm the opposite side of that, when I tried LLMs in coding, it just made my job worse and weird bugs kept sneaking in, of the type that would not happen if I wrote it myself. Also, code review of something that's inherently incapable of learning is a special kind of torture.

And if one ends up writing so much boilerplate, it's a code smell signalling that you need to move stuff to a library.

That said, my main problems with LLMs are ethical. Even if it did its job well, I would not use them because of all the externalities.

This entry was edited (4 days ago)
in reply to Daniël Franke

glitchsoc - Link to source

Dmytri

what are your ethical concerns with free, small, specialized models? I definitely understand the concerns around AI in general, including the ethical dimensions, but I'm quite happy to have simple conversational interface that does basic IDE type stuff, in part, because I'm a GUI hater, and so never liked loads of tool bars and buttons, but using a little conversational interface in my vim terminal buffer to grab a boilerplate k8s manifest, etc, is appealing.
This entry was edited (4 days ago)
⇧